How Does It Work?

Boilerplate

attrs certainly isn’t the first library that aims to simplify class definition in Python. But its declarative approach combined with no runtime overhead lets it stand out.

Once you apply the @attr.s decorator to a class, attrs searches the class object for instances of attr.ibs. Internally they’re a representation of the data passed into attr.ib along with a counter to preserve the order of the attributes.

In order to ensure that subclassing works as you’d expect it to work, attrs also walks the class hierarchy and collects the attributes of all base classes. Please note that attrs does not call super() ever. It will write dunder methods to work on all of those attributes which also has performance benefits due to fewer function calls.

Once attrs knows what attributes it has to work on, it writes the requested dunder methods and – depending on whether you wish to have a dict or slotted class – creates a new class for you (slots=True) or attaches them to the original class (slots=False). While creating new classes is more elegant, we’ve run into several edge cases surrounding metaclasses that make it impossible to go this route unconditionally.

To be very clear: if you define a class with a single attribute without a default value, the generated __init__ will look exactly how you’d expect:

>>> import attr, inspect
>>> @attr.s
... class C(object):
...     x = attr.ib()
>>> print(inspect.getsource(C.__init__))
def __init__(self, x):
    self.x = x

No magic, no meta programming, no expensive introspection at runtime.


Everything until this point happens exactly once when the class is defined. As soon as a class is done, it’s done. And it’s just a regular Python class like any other, except for a single __attrs_attrs__ attribute that can be used for introspection or for writing your own tools and decorators on top of attrs (like attr.asdict()).

And once you start instantiating your classes, attrs is out of your way completely.

This static approach was very much a design goal of attrs and what I strongly believe makes it distinct.

Immutability

In order to give you immutability, attrs will attach a __setattr__ method to your class that raises a attr.exceptions.FrozenInstanceError whenever anyone tries to set an attribute.

Depending on whether a class is a dict class or a slots class, attrs uses a different technique to circumvent that limitation in the __init__ method.

Once constructed, frozen instances don’t differ in any way from regular ones except that you cannot change its attributes.

Dict Classes

Dict classes – i.e. regular classes – simply assign the value directly into the class’ eponymous __dict__ (and there’s nothing we can do to stop the user to do the same).

The performance impact is negligible.

Slots Classes

Slots classes are more complicated. Here it uses (an aggressively cached) object.__setattr__() to set your attributes. This is (still) slower than a plain assignment:

$ pyperf timeit --rigorous \
      -s "import attr; C = attr.make_class('C', ['x', 'y', 'z'], slots=True)" \
      "C(1, 2, 3)"
........................................
Median +- std dev: 378 ns +- 12 ns

$ pyperf timeit --rigorous \
      -s "import attr; C = attr.make_class('C', ['x', 'y', 'z'], slots=True, frozen=True)" \
      "C(1, 2, 3)"
........................................
Median +- std dev: 676 ns +- 16 ns

So on a standard notebook the difference is about 300 nanoseconds (1 second is 1,000,000,000 nanoseconds). It’s certainly something you’ll feel in a hot loop but shouldn’t matter in normal code. Pick what’s more important to you.

Summary

You should avoid instantiating lots of frozen slotted classes (i.e. @attr.s(slots=True, frozen=True)) in performance-critical code.

Frozen dict classes have barely a performance impact, unfrozen slotted classes are even faster than unfrozen dict classes (i.e. regular classes).